REVIEW 3 major objections 6 minor 13 references
CNN-Based Segmentation of the Cardiac Chambers and Great Vessels in Non-Contrast-Enhanced Cardiac CT
T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read A network trained on virtual non-contrast CT images can segment the heart's chambers and large vessels in real non-contrast cardiac CT, enabling volume measurement without contrast injection.
desk verdict Clever VNC-to-NCCT domain trick with solid VNC validation, but the headline NCCT volume claim rests on expert grades, not measurements. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the virtual non-contrast (VNC) image, reconstructed by a dual-layer detector CT scanner from a contrast-enhanced acquisition. Since CCTA and VNC come from the same scan, they are perfectly co-registered, so reference segmentations drawn on CCTA transfer directly to VNC images, eliminating the need for manual NCCT annotation. The segmentation model is a 2D residual fully convolutional network that takes 256×256×5 voxel inputs and outputs per-class segmentation maps, trained by minimizing the sum of soft Dice losses over all seven structures.
What would settle it
Acquire a cohort of patients who receive both a true non-contrast cardiac CT and, on the same or nearly the same day, a contrast-enhanced dual-layer CT from which VNC images can be reconstructed. Train the network exactly as described, then compare volumes measured from the real NCCT segmentations against volumes from the paired CCTA or a cardiac MRI reference. If the paired volume differences exceed the inter-observer variability of the reference method, or if Dice on real NCCT with a quantitative reference is much lower than the VNC Dice, the transfer assumption fails.
Extended reading notes
Core claim
The central discovery is that the domain gap between contrast-enhanced and non-contrast cardiac CT can be bridged using the physics of dual-layer detector CT rather than by manual annotation or image registration. Because a VNC image is reconstructed from the same acquisition as the coronary CT angiography (CCTA) image, the two are perfectly aligned, so pixel-level reference labels drawn on CCTA are valid for the VNC image at zero labeling cost. A fully convolutional network trained on VNC images with CCTA labels segments seven structures—left ventricular cavity and myocardium, right ventricle, left and right atria, ascending aorta, and pulmonary artery trunk—in both VNC and true non-contrast CT images. The paper reports mean Dice scores from 0.84 to 0.94 on VNC test folds and qualitative expert grading of 218 non-contrast scans, of which 67% received grades 1–2 (very accurate or minor errors) and 7% were judged failed.
Load-bearing premise
The method assumes that virtual non-contrast images reconstructed from a contrast-enhanced dual-layer CT scan are similar enough to true non-contrast CT images that a network trained on the virtual ones will segment the real ones correctly; the paper tests this only with subjective expert grading on 218 real scans, not with quantitative volume validation.
Editorial extensions
If this is right
- Cardiac chamber and great-vessel volumes can be estimated from non-contrast CT without manual annotation, extending volumetric risk assessment to patients who never receive contrast.
- A network trained once on dual-layer VNC data segments NCCT images from a different scanner, suggesting the approach does not require the training scanner to be able to produce VNC images at inference time.
- The same acquisition supplies both the training image and the reference label, removing the need for inter-modality registration between CCTA and NCCT.
- The reported Dice scores and the expert grading of 218 real NCCT scans indicate that the segmentations are accurate enough for automated volume measurement in practice.
Reading between the lines
- Beyond the paper: if the VNC-to-NCCT transfer is driven by image appearance rather than by the specific scanner, the same training recipe should work for other dual-layer scanners and for other anatomies, though the paper's evidence is limited to seven cardiac structures and one scanner family.
- Beyond the paper: a direct quantitative check would segment paired NCCT and CCTA images from the same patients and compare volumes; agreement within clinical tolerances would validate the subjective grades.
- Beyond the paper: the residual grade 3–5 failures could be reduced by adversarial domain adaptation from the VNC domain to the NCCT domain, a direction the paper itself names as future work.
- Beyond the paper: the strategy inverts the annotation burden—generate a perfectly aligned surrogate of the target domain instead of labeling the target domain—so it could be reused whenever a physics-based reconstruction aligns two images of the same anatomy.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This extended abstract proposes a fully convolutional network (FCN) for segmenting seven cardiac structures in non-contrast CT (NCCT) images. Since manual reference segmentations in NCCT are hard to obtain, the authors train the FCN on virtual non-contrast (VNC) images reconstructed from dual-layer detector CT acquisitions, using reference segmentations derived from perfectly aligned CCTA images. The model is evaluated on 18 VNC images via nested cross-validation with Dice and ASSD metrics, and on 218 NCCT images from a different scanner via qualitative expert grading. The authors conclude that the FCN enables accurate volume quantification of cardiac chambers and great vessels in the absence of contrast injection.
Significance. If the claimed transfer from VNC to NCCT holds, the method is practically valuable: it leverages the physical alignment of dual-layer CT to avoid manual NCCT annotation, and the reported VNC Dice scores (0.84–0.94) are competitive. The nested cross-validation design on the 18 VNC images is methodologically sound and prevents label leakage, and the ensemble of six fold-models is a sensible testing strategy. However, the clinical volume-quantification claim rests entirely on the generalization from VNC to real NCCT, and that generalization is currently supported only by subjective expert grading without a reference standard, quantitative error metrics, or baseline comparison. Strengths of the paper are the clean use of a paired CCTA/VNC acquisition and the explicit reporting of per-structure Dice and ASSD on the primary set; the weakness is the lack of any quantitative validation on the secondary NCCT set.
major comments (3)
- [Section 3 (Experiments and results), secondary data set evaluation] The central claim that the FCN enables accurate volume quantification in NCCT is supported only by expert grading (Grades 1–5) of segmentations on the 218 NCCT images. This evaluation has no reference standard, no quantitative metric such as Dice or ASSD, no confidence intervals, and no inter-observer variability. The sentence 'most segmentations contained slight errors that are unlikely to significantly impact volume measurements' is an assertion, not a measurement; visually acceptable segmentations can still be systematically biased, for example by consistent over- or under-segmentation of the LV myocardium or atrial walls at low contrast, which would directly invalidate volume estimates. A quantitative evaluation, such as comparing NCCT-derived volumes with CCTA-derived volumes in the same patients or manual annotation on a random subset, is needed to support the volume-quantification claim.
- [Section 2 (Materials and methods), VNC-to-NCCT domain transfer] The method's premise is that VNC images mimic real NCCT images, but this premise is never quantitatively validated. The primary-set Dice/ASSD measurements are all on VNC images from a Philips IQon scanner, while the secondary NCCT images are from a different scanner (Philips Brilliance iCT) and reconstruction without VNC capability. No comparison of image-intensity distributions, noise properties, or anatomical boundary appearance is provided, and no baseline comparison is included (e.g., an FCN trained directly on CCTA images, or the multi-atlas method of Shahzad et al. (2017)). Because the entire training strategy depends on VNC as a proxy for NCCT, this missing quantitative transfer analysis is load-bearing for the paper's main conclusion.
- [Section 4 (Discussion and conclusion) and Section 3] The conclusion that 'this allows accurate volume quantification' is not directly tested anywhere in the manuscript. Volumes are never computed from either the VNC or the NCCT segmentations, and no comparison is made to reference volumes from CCTA or any other modality. Since the secondary data set includes corresponding CCTA images for each NCCT scan, a volume comparison (e.g., segmenting the CCTA images with an existing method and correlating volumes, or measuring bias via Bland-Altman analysis) is feasible and would provide a concrete test of the central claim. Without such a test, the paper's clinical conclusion overreaches the evidence presented.
minor comments (6)
- [Section 3] The grading criteria of Abadi et al. (2010) are referenced but not described; a brief definition of Grades 1–5 is needed for the reader to interpret the reported distribution (12% Grade 1, 55% Grade 2, 19% Grade 3, 7% Grade 4, 7% Grade 5).
- [Section 2] The manuscript says images were smoothed with a 'moderate Gaussian filter' but does not specify the kernel size or standard deviation; please provide these parameters for reproducibility.
- [Section 2] The text says the architecture is based on the '2D residual FCN' by Johnson et al. (2016) but then describes 256×256×5 voxel 3D inputs; it is unclear whether 3D convolutions were used or whether slices were processed independently. This should be clarified.
- [Table 1] The table caption should state explicitly that the results are for 18 VNC images from the primary data set, since the sample size is not mentioned in the caption.
- [Figure 2] Figure 2 shows a single NCCT segmentation example; given the wide range of expert grades, a montage illustrating examples of each grade would be more informative.
- [General] The paper does not state whether the study had institutional review board approval or whether data are available; for a clinical imaging study this information should be included, even in an extended abstract.
Circularity Check
No significant circularity: the method is an empirical training and evaluation procedure with independent train/test separation.
full rationale
The paper does not derive a prediction from a fitted parameter or from an imported uniqueness theorem. The method trains an FCN on virtual non-contrast (VNC) images with reference segmentations obtained on perfectly aligned CCTA images, and evaluates it with six-fold nested cross-validation on the 18 VNC images. The secondary evaluation on 218 NCCT images is independent of the training data, and the expert grading is an external, albeit qualitative, assessment. The only load-bearing assumption, that VNC images sufficiently mimic real NCCT images for domain transfer, is a stated premise rather than a conclusion derived from itself; it is not supported by the paper's own outputs, but that is an evidence-strength concern, not circularity. Self-citation of van Hamersvelt et al. (2019) only describes the primary data set and does not supply a result used to justify the central claim. No fitted constant is renamed as a prediction, and no equation reduces to its own input. The paper is therefore self-contained as an empirical study, and the circularity score is 0.
Assumptions & free parameters
free parameters (7)
- Isotropic resampling voxel size =
0.8 mm
- Input patch size =
256x256x5 voxels
- Number of down/upsampling layers =
3
- Number of residual blocks =
6
- Mini-batch size =
32
- Initial learning rate and decay schedule =
0.001, 70% decay every 2000 iterations
- Gaussian smoothing filter =
not specified ('moderate')
assumptions (4)
- domain assumption VNC images mimic real NCCT images
- domain assumption CCTA and VNC images from the same dual-layer acquisition are perfectly aligned
- domain assumption Reference segmentations obtained on CCTA are valid for VNC images
- domain assumption Expert visual grading with Abadi criteria is a valid proxy for segmentation quality and volume accuracy
Cite this review
Pith. "Pith review of CNN-Based Segmentation of the Cardiac Chambers and Great Vessels in Non-Contrast-Enhanced Cardiac CT." pith.science (2026). https://pith.science/paper/QNWBSQDK
@misc{pith2026190807727,
author = {Pith},
title = {Pith review of: CNN-Based Segmentation of the Cardiac Chambers and Great Vessels in Non-Contrast-Enhanced Cardiac CT},
year = {2026},
howpublished = {\url{https://pith.science/paper/QNWBSQDK}},
note = {Machine review of arXiv:1908.07727}
}
read the original abstract
Quantification of cardiac structures in non-contrast CT (NCCT) could improve cardiovascular risk stratification. However, setting a manual reference to train a fully convolutional network (FCN) for automatic segmentation of NCCT images is hardly feasible, and an FCN trained on coronary CT angiography (CCTA) images would not generalize to NCCT. Therefore, we propose to train an FCN with virtual non-contrast (VNC) images from a dual-layer detector CT scanner and a reference standard obtained on perfectly aligned CCTA images.
Figures
Reference graph
Works this paper leans on
-
[1]
Sobhi Abadi, Ariel Roguin, Ahuva Engel, and Jonathan Lessick. Feasibility of automatic assessment of four-chamber cardiac function with MDCT: initial clinical application and validation. European Journal of Radiology, 74 0 (1): 0 175--181, 2010
work page 2010
-
[2]
Automatic model-based segmentation of the heart in CT images
Olivier Ecabert, Jochen Peters, Hauke Schramm, Cristian Lorenz, Jens von Berg, Matthew J Walker, Mani Vembar, Mark E Olszewski, Krishna Subramanyan, Guy Lavi, et al. Automatic model-based segmentation of the heart in CT images. IEEE Transactions on Medical Imaging, 27 0 (9): 0 1189, 2008
work page 2008
-
[3]
Perceptual losses for real-time style transfer and super-resolution
Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694--711. Springer, 2016
work page 2016
-
[4]
Hortense A Kiri s li, Michiel Schaap, Stefan Klein, Stella-Lida Papadopoulou, Mara Bonardi, Chin-Hui Chen, Annick C Weustink, Nico R Mollet, Evert-Jan Vonken, Rob J van der Geest, et al. Evaluation of a multi-atlas based method for segmentation of cardiac CTA data: a large-scale, multicenter, and multivendor study. Medical Physics, 37 0 (12): 0 6279--6291, 2010
work page 2010
-
[5]
Domain-adversarial neural networks to address the appearance variability of histopathology images
Maxime W Lafarge, Josien PW Pluim, Koen AJ Eppenhof, Pim Moeskops, and Mitko Veta. Domain-adversarial neural networks to address the appearance variability of histopathology images. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support, pages 83--91. Springer, 2017
work page 2017
-
[6]
Harsha S Nagarajarao, Alan D Penman, Herman A Taylor, Thomas H Mosley, Kenneth Butler, Thomas N Skelton, Tandaw E Samdarshi, Giorgio Aru, and Ervin R Fox. The predictive value of left atrial size for incident ischemic stroke and all-cause mortality in African Americans : the atherosclerosis risk in communities ( ARIC ) study. Stroke, 39 0 (10): 0 2701--2706, 2008
work page 2008
-
[7]
Left ventricular diameter and risk stratification for sudden cardiac death
Kumar Narayanan, Kyndaron Reinier, Carmen Teodorescu, Audrey Uy-Evanado, Ryan Aleong, Harpriya Chugh, Gregory A Nichols, Karen Gunson, Barry London, Jonathan Jui, et al. Left ventricular diameter and risk stratification for sudden cardiac death. Journal of the American Heart Association, 3 0 (5): 0 e001193, 2014
work page 2014
-
[8]
Subha V Raman, Mona Shah, Beth McCarthy, Anne Garcia, and Amy K Ferketich. Multi--detector row cardiac computed tomography accurately quantifies right and left ventricular size and function compared with cardiac magnetic resonance. American Heart Journal, 151 0 (3): 0 736--744, 2006
work page 2006
Show all 13 references
-
[9]
CT calcium scoring
Veit Sandfort and David A Bluemke. CT calcium scoring. History , current status and outlook. Diagnostic and Interventional Imaging, 98 0 (1): 0 3--10, 2017
2017
-
[10]
Automatic segmentation and quantification of the cardiac structures from non-contrast-enhanced cardiac CT scans
Rahil Shahzad, Daniel Bos, Ricardo PJ Budde, Karlijn Pellikaan, Wiro J Niessen, Aad van der Lugt, and Theo van Walsum. Automatic segmentation and quantification of the cardiac structures from non-contrast-enhanced cardiac CT scans. Physics in Medicine & Biology, 62 0 (9): 0 3798, 2017
2017
-
[11]
Application of spe C tra L computed tomogr A phy to imp R ove spec I ficity of cardiac compu T ed tomograph Y (clarity study): rationale and design
Robbert W van Hamersvelt, Ivana I s gum, Pim A de Jong, Maarten JM Cramer, Geert EH Leenders, Martin J Willemink, Michiel Voskuil, and Tim Leiner. Application of spe C tra L computed tomogr A phy to imp R ove spec I ficity of cardiac compu T ed tomograph Y (clarity study): rat...
2019
-
[12]
Four-chamber heart modeling and automatic segmentation for 3-D cardiac CT volumes using marginal space learning and steerable features
Yefeng Zheng, Adrian Barbu, Bogdan Georgescu, Michael Scheuering, and Dorin Comaniciu. Four-chamber heart modeling and automatic segmentation for 3-D cardiac CT volumes using marginal space learning and steerable features. IEEE Transactions on Medical Imaging, 27 0 (11): 0 166...
2008
-
[13]
Evaluation of algorithms for multi-modality whole heart segmentation: An open-access grand challenge
Xiahai Zhuang, Lei Li, Christian Payer, Darko Stern, Martin Urschler, Mattias P Heinrich, Julien Oster, Chunliang Wang, Orjan Smedby, Cheng Bian, Xin Yan, Pheng-Ann Heng, Aliasghar Mortazi, Ulas Bagci, Guanyu Yang, Chenchen Sun, Gaetan Galisot, Jean-Yves Ramel, Thierry Brouard...
1902 arXiv
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.